Copied


NVIDIA Launches CUDA Rust for GPU Kernels, Expands Rust Ecosystem

James Ding   Sep 08, 2026 13:26 0 Min Read


NVIDIA has officially unveiled CUDA Rust, opening a new chapter for GPU programming by enabling developers to write GPU kernels directly in Rust. Announced on September 8, 2026, through the NVIDIA Technical Blog, the initiative introduces two distinct programming tracks—cuda-oxide and cutile-rs—targeting both advanced and general-purpose use cases.

CUDA Rust addresses a long-standing gap in Rust's GPU ecosystem. Historically, Rust has served as a wrapper for GPU kernels written in other languages like CUDA C++. CUDA Rust changes this paradigm by allowing developers to compile native Rust kernels directly to PTX (Parallel Thread Execution), NVIDIA's low-level assembly language for GPUs. This move aligns with the industry's growing adoption of Rust for systems programming, thanks to its unique ability to catch bugs at compile time without sacrificing performance.

Two Tracks: SIMT and Tile

The cuda-oxide track employs a SIMT (single instruction, multiple threads) model, mirroring traditional CUDA C++ workflows. It uses a custom rustc codegen backend to compile Rust kernels to PTX. Meanwhile, the cutile-rs track leverages an emerging Tile programming model, which abstracts GPU thread management by focusing on tiles (sub-tensors of data) rather than individual threads. Unlike cuda-oxide, cutile-rs operates on stable Rust and is already in use by industry players like Hugging Face and mistral.rs.

For developers, the choice between cuda-oxide and cutile-rs depends on specific project needs. cuda-oxide offers fine-grained control, ideal for optimizing memory and thread usage, while cutile-rs simplifies development with a safer, higher-level abstraction. Both tracks promise interoperability with CUDA C++ and CUDA Python, ensuring developers are not locked into a single language stack.

Rust Ecosystem Gains Traction

NVIDIA's announcement comes as Rust gains momentum in GPU programming. In May 2026, the Rust team announced a major update in version 1.97, raising support for NVIDIA GPU targets to PTX ISA 7.0 and SM 7.0. NVIDIA itself has already integrated Rust into core projects like its Nova Linux driver and Dynamo AI system. The CUDA Rust launch further underscores the company's commitment to Rust as a foundational language for its software stack.

Early adoption metrics suggest CUDA Rust could become a significant player in the GPU programming ecosystem. NVIDIA Research highlighted cutile-rs in its 2026 study, “Fearless Concurrency on the GPU,” which demonstrated robust performance for element-wise operations and GEMM workloads on NVIDIA's B200 GPUs.

Market Context

The launch of CUDA Rust also coincides with NVIDIA's strong market performance. As of September 8, 2026, NVIDIA's stock price (NVDA) stands at $230.36, marking a 0.77% gain over the past 24 hours. With a market cap of $5.59 trillion, NVIDIA remains a dominant player in the AI and GPU sectors. CUDA Rust could further solidify its position by attracting Rust developers to NVIDIA's ecosystem and expanding the toolkit for AI and high-performance computing applications.

What’s Next?

Both cuda-oxide and cutile-rs are early-stage projects and not yet ready for production use. Developers eager to experiment can find documentation and examples via the cuda-oxide and cutile-rs GitHub repositories. NVIDIA is also soliciting feedback from the community to refine these tools and address usability gaps.

For those attending RustConf 2026 in Montréal, NVIDIA engineer Melih Elibol will present “Fearless Concurrency on the GPU,” detailing CUDA Rust's capabilities and roadmap. This marks an opportunity for Rust and GPU developers to engage directly with NVIDIA's team and help shape the future of Rust on GPUs.

In the near term, CUDA Rust's growth will depend on its adoption by the Rust community and its ability to bridge the gap between high-level programming safety and low-level GPU performance. With NVIDIA's engineering muscle behind it, CUDA Rust has the potential to redefine how developers approach GPU programming in the years to come.


Read More